Abstract
Deep learning (DL) models have shown strong performance in Prognostics and Health Management (PHM) of cyber-physical systems (CPS), but their high computational cost and limited interpretability pose a critical barrier to deployment in real-time, resource-constrained industrial environments. This limitation is particularly restrictive for PHM applications, where models must operate under strict latency and hardware constraints. This paper studies these challenges jointly by combining model compression with structured explainability. A transformer-based fault detection model is evaluated using pruning, quantization, and knowledge distillation to explicitly reduce computational and memory requirements while preserving diagnostic performance. In parallel, a SHAP-based explanation module is developed to organize feature attributions into operator-oriented indicators, including dominant sensors, deviation direction, and affected plant stages. The approach is evaluated on publicly available Secure Water Treatment (SWaT) datasets. Results show that quantization-aware training and knowledge distillation achieve favorable efficiency-performance trade-offs, while the proposed explanation framework produces consistent and focused diagnostic cues across attack scenarios. The study provides insight into the interaction between efficiency and interpretability and supports the development of lightweight, deployable, and explainable PHM models for cyber-physical systems.